rotating-anomaly-detector

1-D convolutional autoencoder for unsupervised anomaly detection on rolling-bearing vibration signals. Trained on healthy-only data; windows the model cannot reconstruct (high reconstruction MSE) are flagged as anomalous.

Model

ConvAutoencoder (PyTorch): 3 × (Conv1d stride-2 + ReLU) encoder → 32-d latent vector → mirrored transposed-convolution decoder. ~1.1M parameters (see model.py). Input: 2048-sample windows, per-window z-normalised.

Results (reproduced locally on CPU)

Trained on NORMAL-ONLY windows (CWRU 98.mat, drive-end, first 80% of the signal, stride-512 windows → 753 windows; 50 epochs, Adam 1e-3, ~13 min CPU). The alarm threshold (0.3397) is the 95th percentile of the normal validation scores — fixed before any fault window was seen. Tested on held-out normal windows (23) + all 9 fault files (2081 windows).

Metric	Value
AUROC (normal vs fault windows)	1.000
Precision @ threshold	0.9995
Recall @ threshold	1.000
F1 @ threshold	0.9998

Per fault condition, recall at the fixed threshold: 100% for all nine (ball / inner-race / outer-race × 0.007" / 0.014" / 0.021"). Mean fault scores are 2.8–3.3× the normal mean. The one imperfection: 1 of 23 healthy test windows trips the threshold (≈4% false-alarm rate, as designed by the 95th-percentile rule).

Usage

from model import load_model, anomaly_score

model = load_model("pytorch_model.bin", device="cpu")
scores = anomaly_score(model, signal)   # per-window reconstruction errors
threshold = 0.3397                     # from config.json
flagged = scores >= threshold
print(f"{100*flagged.mean():.1f}% of windows above threshold")

See config.json for hyperparameters, the threshold, and the exact preprocessing/train protocol.

Dataset

Case Western Reserve University Bearing Data Center, 48 kHz drive-end recordings (1 HP load, ~1772 rpm). Ten conditions: healthy (98.mat) plus seeded ball / inner-race / outer-race faults at three severities.

Limitations (read before deploying)

• CWRU faults are seeded (EDM-drilled) defects on a single test rig, one load, one speed — not naturally evolving field faults.
• Trained on one operating condition; a speed/load change shifts the vibration signature and needs re-validation (or domain adaptation).
• The 95th-percentile rule bakes in ≈5% false alarms on healthy data; in production add a persistence filter (e.g. alarm only after K consecutive flagged windows).
• Threshold estimated from a small normal validation set (23 windows).